LoGRA cuts RL training memory by 45.7%, enabling 27B RL on a single 8-GPU node
青稞AI · wechat · 2026-10-12
LoGRA tackles the memory bottleneck of RL fine-tuning large models. Instead of freezing the base model and training a LoRA, it accumulates low-rank gradients directly during backprop without storing full gradient matrices, rotating random projections each step while writing updates back to the original model — avoiding the fixed low-rank subspace constraint. A predicted-KL estimate modulates step size to prevent overshooting policy updates.
Results: up to 45.7% training memory savings; RL on a 27B model runs on a single 8-GPU node with 1,100+ stable steps where standard Adam OOMs. Code open-sourced on top of Molt; paper at arXiv 2610.06647.
Related event: NVIDIA's LoGRA Cuts RL Training Memory by 45.7%(2 posts)→
More from Infra
- NVIDIA Vera Rubin Ultra NVL576 rack slide shown, built for 10T+ parameter models — firstadopter · 2026-10-13
- Over 1m gaming and workstation chips repurposed for AI as China's Nvidia gap persists to 2030 — zijing_wu · 2026-10-13
- Lizard converts pretrained Transformers to subquadratic architectures with near-lossless accuracy — burkov · 2026-10-13
- NVIDIA unveils first 800 VDC AI Factory design with 16 MW deployment units — firstadopter · 2026-10-13
- Broadcom backs hybrid copper-plus-optics racks to cut scale-up optics costs — firstadopter · 2026-10-13
- Blog post shrinks from 158KB HTML to 17KB Markdown for AI agents — colinmcnamara · 2026-10-13